Does GPT-4 Decode Richer Dimensional Structure of Emotion? Mapping Dimensional Structure with 99 Emotions

Dimensional theories of emotion propose that affective experiences can be organized along continuous dimensions like Pleasure, Arousal, and Dominance. Recent computational methods for capturing these conceptual structures have relied primarily on word embeddings, requiring substantial technical expertise. This study examined whether prompting GPT-4 as a more accessible method, can capture this structure more effectively, by comparing both approaches against human spatial reasoning, with particular focus on analyzing extensive emotion vocabularies exceeding constraints of human cognitive capability. Study 1 established validity by testing six basic emotions, prompting showed strong convergence with human judgments, while embeddings produced fundamentally different clustering patterns. Study 2 extended analysis to 99 emotion terms. Both approaches identified two optimal clusters distinguished by both Pleasure and Dominance, though prompting produced clearer cluster separation. Particularly, Arousal's discriminative power increased substantially in case of 99 emotion terms with prompting method only. This pattern may reflect how broader sampling of emotion lexicon provides sufficient linguistic contexts for arousal-related distinctions to emerge statistically. Together these results indicated that GPT prompting offers a promising methodological tool for dimensional emotion research, particularly in terms of uncovering patterns that only become visible when analysing a broader sample of emotion lexicon.